Despite the fact that we spend roughly a third of our lives snoozing (or at least trying to), sleep is not well understood by scientists — to say nothing of the estimated 35% of Americans who don’t get enough of it.
Part of the problem is how difficult it is to study slumber. Experiments currently require individuals to come into a hospital or laboratory, cover themselves with an army of electrodes and allow a team of doctors to observe their attempts to drift off. “It’s kind of a nightmare,” says Dr. Emmanuel Mignot, director of the Stanford Center for Sleep Sciences and Medicine. “They’re hooked up to all these things and they can barely move.” This environment can make it challenging to capture data about normal sleep and is often prohibitively expensive to maintain for longer than a night or two — a necessity for researchers looking to understand trends over time.
“When you are a patient, when you have problems, the last thing you want to do is add to those problems by having to wear sensors, or take measurements or write diaries about how you feel,” says Dina Katabi, a professor of electrical engineering and computer science at MIT. “If we can monitor health continuously but passively in a patient’s natural environment, that can help dramatically.”
Experts say artificial intelligence (AI) might help researchers do just that. “It has the potential of changing sleep medicine entirely,” says Mignot. “Sleep is going through a revolution.”
In 2017, Katabi and her team tested a way to monitor sleep information without any wires or electrodes. Instead of FDA-approved sleep staging devices, they used a device that emits radio frequency signals that bounce off the body. Because even tiny movements — think muscular twitches or blood pulsing, not rolling over or kicking a leg — alter the way these signals travel, the researchers developed an AI algorithm that would associate the data obtained from the radio waves with certain stages of sleep. In a trial monitoring 25 people, they found that the AI system could accurately guess an individual’s sleep stage about 80% of the time using only radio signal data, providing valuable information about their overall rest quality and problems they might be experiencing. The device is currently available for research purposes, but it is not sold commercially.
Katabi allows that some accuracy is lost without direct brain-activity monitoring, but she says the system offers huge advantages over traditional sleep studies, given how accessible and non-invasive it is — perks that could allow researchers to collect a higher volume of sleep data without inconveniencing patients.
Once this data is collected, however, it still needs to be analyzed and interpreted. And here, too, AI could make an impact.
Right now, sleep researchers manually pore over pages of sensor-generated data related to brain activity, eye movements, breathing patterns, leg kicks and more, and use it to make assessments about sleep quality and possible problems, such as sleep apnea and narcolepsy. But in a 2018 paper published in Nature Communications, Stanford’s Mignot and his colleagues showed that an AI system could use this data to detect sleep issues better than a human technician.
First, they asked six technicians to examine and score a single set of sleep data, looking for abnormalities that could suggest narcolepsy. The researchers then averaged those scores to get a consensus value. Next, the researchers trained an AI system by showing it 3,000 sleep readings, so it could begin to associate certain conditions with specific data trends. Then they had the system score the same sleep data as the technicians. It got closer to the group’s consensus value than any single human did, suggesting that it was more accurate than the average technician.
“You are doing better than the standard of care, and it’s cheaper and it’s faster and it’s more reproducible,” Mignot says. While the innovation has not yet been approved for clinical use, Mignot says the science is sound.
But even as scientists like Mignot and Katabi are carefully studying AI and verifying their results with rigorous tests, consumer-facing companies are already jumping into bed with the technology.
This month, sleep tech company BRYTE introduced a smart mattress that learns from data measured by internal sensors as well as feedback provided by the user until it can improve sleep by adjusting its temperature, support and surrounding light in real-time. Mattress company Eight also syncs its smart mattresses with an AI-powered sleep coach app that delivers sleep insights and recommendations based on the person’s sleep history; the bed can also be hooked up to other smart home features, such as lights and thermostats, to adjust environmental conditions automatically.
Outside the mattress realm, the Apple Watch says it can automatically track sleep quality if users wear it to bed and uses the information it gathers to make personalized recommendations for better rest. And apps like Sonic Sleep (which uses a phone’s microphone to assess background noise and then employs an AI-powered algorithm to pick out the right audio to mask it) and Sleep.ai (which says it can detect sleep apnea, snoring and teeth grinding by analyzing sound samples) have also gotten in on the AI sleep space.
Perhaps unsurprisingly for any parent who’s ever wondered why their baby won’t sleep soundly through the night, both researchers and brands are also interested in bringing AI’s power to the crib. Apps including Nod and Huckleberry ask parents to log detailed information about their kids’ sleep, and over time use AI to tailor recommendations on how to improve it, like adjusting bedtimes or feeding schedules. And Nanit’s forthcoming motion-sensing infant swaddle, which works with the brand’s HD-camera-equipped baby monitor to automatically track a child’s sleep, provides parents with AI-powered recommendations on how to improve it. The AI for Good Foundation also recently launched a research program aimed at gathering large amounts of sleep data from parents and their children, which will then be analyzed by AI algorithms to provide deeper insights into the connections between good sleep and good health.
Right now, products and projects like these are primarily aimed at accomplishing a straightforward, if elusive, goal: Helping people get better sleep. But both Mignot and Katabi say the relationship between AI and sleep goes far deeper than simple shut-eye.
“Sleep is very important,” Katabi says, “but sleep is not just sleep.”
Precursors and symptoms of a range of chronic conditions manifest while we’re sleeping. Individuals with depression, for example, often also suffer from insomnia. And many people with Parkinson’s disease develop REM sleep behavior disorder — meaning they physically act out their dreams — decades before the emergence of telltale symptoms like tremors. Having access to a tool that could quickly and accurately pick up on these sleep disturbances could help treatment options, Mignot says.
“Many of these diseases we need to be preventing; we can’t wait until the brain is gone,” Mignot says. “Sleep is the window to brain health in many, many cases.”